Am 12.04.2012 23:24, schrieb Anthony Liguori: > I think you're trying to run before we walk.
The opposite is the case: I am trying to walk by improving code that was not even qdev'ified before I touched it (see the sh4 commit in my previous message). But now you are telling me I can't get a function for that because Paolo and you want to run to a stage where realize doesn't create new objects, which in reality today it does. So what about applying a modified version of my patch that clearly marks it as "obsolete"? Then we can at least move forward with sh4 (and prep and ...). There's still quite some state around that just g_malloc()s some struct and passes it around in opaques - that's also where I was wishing for a QOM board object so that board state can be properly modelled. I really don't see the point why I should uselessly make objects SysBus without needing MemoryRegions or GPIO qemu_irqs and in the knowledge that SysBus as an articial non-bus is supposed to die in QOM, just so that I can use the existing qdev_init_nofail() for two-stage construction. initfn is unable to handle errors btw, which is another reason to do object creations in a second-stage constructor. Maybe we need to decouple that from realize altogether and move to three-stage constructors instead? With ObjectClass::realize() being for new and reviewed code, like you just suggested, and simply moving DeviceClass::init() to ObjectClass::init() in addition to .instance_init? Andreas -- SUSE LINUX Products GmbH, Maxfeldstr. 5, 90409 Nürnberg, Germany GF: Jeff Hawn, Jennifer Guild, Felix Imendörffer; HRB 16746 AG Nürnberg
